2019 IEEE 32nd International Symposium on Computer-Based Medical Systems (CBMS)
Osteoporosis Classification Using Texture Features
Farhan Riaz, Rehan Nemati, Hina Ajmal, Ali Hassan, Ernest Edifor, Raheel Nawaz
Abstract— Assessment of osteoporotic disease from the radio-
graph image is a significant challenge. Texture characteristics
when observed from the naked eye for the bone micro-
architecture of the osteoporotic and healthy cases are visually
very similar making it a challenging classification problem.
To extract the discriminative patterns in all the orientations
and scales simultaneously in this study we have proposed an
approach that is based on a combination of multi resolution
Gabor filters and 1D local binary pattern (1DLBP) features.
Gabor filter are used due to their advantages in yielding
a scale and orientation sensitive analysis whereas LBPs are
useful for quantifying microstructural changes in the images.
Our experiment show that the proposed method shows good
classification results with an overall accuracy of about 72.71%
and outperforms the other methods that have been considered
in this paper.
Index Terms— Texture analysis, Pattern recognition, Gabor
filters, Osteoporosis.
I. I NTRODUCTION
Osteoporosis is an age-related growing problem in the
world [1]. Although an awareness about the concerns for
osteoporosis is increasing day by day, the diagnosis and
prevention methods are still inadequate. Osteoporosis is a
skeletal disease that affects the bone strength and makes
them susceptible to the fractures. This disease is described
by damage of bone mass and the weakening bone mi-
croarchitecture [2]. Thus, bone mass is not the only aspect
of importance, the skeletal factors and micro-architecture
features are also of great significance [3]. Bone Mineral
Density (BMD) is a degree of mineral content present in the
bone and is an exceptional pointer for osteoporosis. BMD is
typically defined by the Dual Energy X-ray Absorptiometry
(DXA) [4]. However, using only BMD is unsatisfactory
for guessing the fracture risk precisely [5]. It has been
verified that evaluating osteoporosis using bone mass is
unclear [6]. Combining the bone microarchitecture features
with BMD coefficients provides an excellent way for prevent-
ing fracture risk. Although there are many other technologies
like Trabecular Bone Score but none of them prove to
be very useful in clinical practice [7]. According to many
studies [8], [9] the texture analysis of bone radiographs is a
promising way to assess and identify this disease effectively.
Many techniques use three dimensional (3D) non-invasive
measures to assess the bone structure. Despite the accuracy
of these methods, these are not very practical due to their
complexity, price, and access to the devices. 2D texture
analysis presents a simpler way to evaluate the bone structure
This work was supported by the National University of Sciences and
Technology, Islamabad, Pakistan and Manchester Metropolitan University,
Manchester, UK
2372-9198/19/$31.00 ©2019 IEEE
DOI 10.1109/CBMS.2019.00119
using radiographs [9]. The 2D texture analysis is an indirect
assessment of 3D micro-architecture [10]. This assists in
extracting the information about the bone structure and helps
in separating the healthy from the osteoporotic cases. Many
studies have been conducted based on two Dimensional
(2D) bone radiographs using several texture analysis meth-
ods [11]. Fractal geometry is based on fractal dimension
to characterize the textures displaying irregular and self-
similarity behaviors. Fractal analysis has been proved as a
suitable tool to quantify the bone structure changes due to the
fractal nature of bone [12]. This analysis can be done using
the fractional Brownian motion (fBm) and its increments,
fractional Gaussian noise (fGn) [13]. Lacunarity analysis has
also been proposed for the osteoporosis detection [14]. Gabor
transform and wavelets have also been used to study the bone
changes [15]. Some investigations have been made using
the morphological factors for bone structure analysis [16].
Hybrid methods combining statistical and structural methods
like Local Binary Pattern (LBP) operator have been proposed
that have the ability to identify the changes within the
trabecular bone [17]. A directional multi-scale analysis has
been performed to study the bone texture characterization
and calcaneus radiographs have been analysed using circular
parametric models [18]. In [19], a study has been conducted
which used special feature selection techniques that take into
account the individual prediction ability of the feature and
the inter-feature redundancy to find the most distinct feature
set.
Fig. 1: Normal (left) and Abnormal (right) X-ray images.
Diagnosis of osteoporosis using bone radiographs presents
a great challenge because the texture exhibited by the
healthy and osteoporotic cases are visually very similar
to each other (Fig. 1). In this paper, we have proposed
a method for the automated diagnosis of osteoporosis in
bone radiographs. Our premise is that the structural changes
in trabecular bone can be defined by textural features, as
575
established in the literature [19]. We present an approach
combining Gabor filters and one Dimensional (1D) LBP
that provides significant discriminating features. Briefly, this
approach is based multi-resolution spatial analysis combining
intensity distribution with spatial information. Instead of
using directly using spatial intensity distribution, multiple
scales and orientations of the Gabor filters followed by 1D
LBP operator are used. The combination of Gabor and 1D
LBP improved the representation power of spatial histogram
and provided features that enhance different shapes of the
structure. Our goal is to take benefit from both the Gabor and
1D LBP to extract an effective descriptor that can improve
the performance for the early stage diagnosis of osteoporosis.
The paper is organized as follows: we discuss the dataset
used (Section II) followed by the description of the proposed
method (Section III). Later, we describe our experimental
results (Section IV) followed by conclusive remarks (Sec-
tion V).

II. M ATERIALS Fig. 2: Radiograph with its region of interest (adapted

from [20].
All the experiments are conducted on Texture Charac-
terization of Bone (TCB) challenge data publicly available
at http://www.univ-orleans.fr/i3mto/data. This data consists
visual cortex [21]. Another significant consideration is that
of healthy and osteoporotic radiograph images. Any experi-
the Gabor filters achieve optimal localization in both fre-
ments involving human beings were approved by the Review
quency and spatial domain [21]. The Gabor filtering of an
Board of the institution that provides the data. The data set
image is represented as
includes the annotated images of 58 healthy and 58 osteo-
porotic subjects. The region of interest (ROI) was marked by
a physician by pointing out the anatomical marks that can be Gθ,σ (x, y) = I(x, y) ∗ ψθ,σ (x, y)
simply recognized on the calcaneus image. This guarantees  +∞  +∞
that the ROI be picked up in the same area and in the same = I(x, y)ψθ,σ (x − k, y − l) (1)
alignment from each bone radiography. ROI was of 2.7x2.7 k=−∞ l=−∞

cm2 positioned in a part of the calcaneus that comprises of where I(•) denotes the input images, ψθ,σ (•) represents
only the bone. The pixel size was 105m. Figure 1 presents the impulsive response of a Gabor filter having an orien-
four sample images. The top images are from osteoporotic tation θ and scale σ whereas ∗ represents the operator of
patients and the bottom two are from the healthy subjects. convolution. A Gabor filter can be mathematically defined
These calcaneus images present a interesting dataset, since as follows:
all of them show similar visual characteristics.  
f2 − f2 2
x2 + fη2 y 2 

III. M ETHODS ψ(x, y) = e γ2 ej2πf x (2)

πγη
The visual illustration of a region of interest in the
radiograph of a heel is shown in Fig. 2. The heel is subjected x = x cos θ + y sin θ
(3)
to two forces, gravity (tension) and walking force (compres- y  = −x sin θ + y cos θ
sion) which create anisotropic properties in the bone struc- where f is the centre frequency of the filter, θ is the
ture. Moreover, there is a different in granularity of the bones orientation of the major axis of a Gaussian function, γ is the
for normal (dense) and osteoporotic bones (scale changes). standard deviation along the major axis and η is standard
Therefore, the multiresolution analysis using Gabor filters deviation along the minor axis. λ = γη represents the aspect
(orientation and scale variations) and microstructural analysis ration of the Gaussian function. A 2D Gabor filter function
(local binary patterns) becomes very interesting. in the frequency domain is written as
A. Gabor filtering −π 2
(γ 2 (u −f )2 +η2 v2 )
Ψ(u, v) = e f2 (4)
The osteoporosis images are subjected to feature extraction
using a combination of Gabor filters and Local Binary
Patterns (LBP). Over the recent years, Gabor filters have u = u cos θ + v sin θ
(5)
been used widely to address various issues related to texture v  = −u sin θ + vcosθ
classification. The usage of Gabor filters is motivated by the which represents a bandpass filter. A bank of Gabor filters
fact that they are very similar with the primary mammalian is obtained by changing the parameters of a Gabor wavelet

576
and obtaining the relevant filter banks. If S is the total an optimum hyperplane that maximizes the margin of sepa-
number of scales and K is the total number of orientations ration between two different classes. This approach typically
at which the Gabor filters are calculated, we obtain a feature constructs the classification models which have excellent
vector of size S×K for every pixel in a Gabor filtered image. generalization ability thus making it a powerful tool in
various applications. For our implementation, we have used
B. 1D Projections of Filtered Responses SVM with linear kernel for classification.
The filtered images are subjected to the calculations of
their projections along the horizontal and vertical axis. As a IV. E XPERIMENTS
result of this projection, we can potentially find out how the
In this section, all the experiments are reported on the TCB
variations are taking place either horizontally or vertically in
challenge data. At first, the Gabor filter bank consisting of
the images. This is a very reasonable strategy to calculate the
filter of four scales and four orientations are constructed.
variations in the images given that the subject changes will
Later, 1DLBP histograms are obtained. For the calculation
be reflected in the response of the relevant filter bank. Given
of LBPs, 8 neighbors are considered at a spacing of 1 pixel.
filter output Gθ,σ (x, y), its horizontal and vertical projections
Feature extraction is followed by feature reduction using
are calculated as
PCA, keeping about 95% variance of the data. Classification
is done using two well known classifiers: the 1 Nearest
Rows
 Cols
 Neighbours (1NN) and Support Vector Machine (SVM).
GH
θ,σ = Gθ,σ (x, y), GVθ,σ = Gθ,σ (x, y)
A ten-fold cross validation is performed to evaluate the
i=0 i=0
performance of our method. In this work we first compared
C. 1D Local Binary Patterns (LBP) our method using LBP and 1DLBP. The mask is an important
The projection vectors are now subjected to 1D LBPs. parameter determining the classification performance. Thus,
The LBPs were first proposed by Ojala et al. [22] and the results for LBP with 8 neighbours and 1DLBP with 8
than later have been widely used for the texture feature neighbours are used.
extraction. The state-of-the-art representation of LBP used We compared our method with the technique proposed
eight neighboring pixels for the neighborhood information in [17]. We replicated the method and compiled the results
and used the center pixel as the threshold value to encode the for all 116 images of the data set for 1NN and SVM. A part
other pixel information. The 2D LBP is not very useful for from this technique that exists in the literature with refer-
bone texture characterization because bone structure mainly ence to osteoporotic image classification, we compared the
exhibits anisotropic changes. The 2D LBP captures only the performance obtained using two other state-of-the-art texture
frequency of the local structures and have no consideration
for their global orientation and thus is less sensitive to
anisotropic changes. An illustration for the basic concept of
1DLBP that we have adopted is given in Fig. 3.
After the calculation of 1DLBPs for every filter bank both
along the horizontal and vertical directions, we calculate the
histograms of LBP which can be used as features for the
images.
D. Dimensionality reduction
For feature reduction, we have performed principal com-
ponent analysis (PCA). The PCA decomposes the data such
that the features are organized in the order of decreasing
variance values of the decomposed components of the signal.
Effectivly, PCA can be used to fit the data on an n-
dimensional space where the value of  n is chosen such that
a specific percentage of variance of the data is incorporated
in the transformed feature space.
E. Classification
We have chosen different statistical classifiers in our work,
their choice is chosen by their main archetypes:
Nearest Neighbor: This classifier labels an unknown
object with the label of the majority class of its ’k’ neighbors
in the feature space [24]. The distance between an object and
its nearest neighbors is measured in the Euclidean sense.
Support Vector Machines: The SVM, originally pro- Fig. 3: The principle of obtaining 1D Local Binary Patterns
posed by Vapnik et al. [25] mainly consists of constructing (adapted from [23]).

577
feature extraction methods i.e., homogeneous texture [26]
and local binary patterns [22]. Our experiments show that the
proporsed method outperforms the other methods that have
been considered in this paper, irrespective of the classifier
used. We suspect that this is because of the multiresolution
and multidirectional nature of the proposed descriptor as
compared to ther other counterparts that have been used. Al-
though the HT is also a multiresolution descriptor, it exhibits
a shortcoming of the aggregation of the filter responses thus
ignoring the details in the images.
Although reasonable results are obtained, it should be
noted that there is a scope for improvement in the results. We
have not adopted the more conventional ways of using the
convolutional neural networks (CNN) in this paper because
we face several challenges, the most important one being that
the dataset pertaining to osteoporosis is limited (only 118
images, half from normal bones and half from osteoporotic
bones).
Method SVM KNN
HT 62.57 64.18
LBP 59.64 60.25
Wavelet-1DLBP 67.94 68.62
Proposed 70.59 72.71
TABLE I: Detection Results for Osteoporosis.
V. C ONCLUSIONS
In this paper, we have proposed a novel technique for
the extraction of texture features from X-Ray radiographs.
The method is based on two state-of-the-art texture feature
extraction methods: Gabor filters and local binary patterns
which exhibit their own distinct advantages. The texture
features are obtained using Gabor filtering first, followed by
the calculation of the projected Gabor responses (along hori-
zontal and vertical directions), obtaining two vectors for each
filtered output. This output is subjected to the calculation of
1DLBPs followed by feature reduction and classification of
the images. It should be noted that the strengths of both LBP
(microstructure analysis) and Gabor filters (multiresolution
analysis) are utilized in the proposed features. The idea is
to capture multi-resolution global and local characteristics
patterns to achieve a better characterization of the images.
Our results show that the proposed feature set shows good
results, although there is still scope for improvement.
There are several issues with this study, which are mainly
triggered by the lack of adequate data. Due to this limitation,
it is not very feasible to use more sophisticated classification
methods such as CNNs on the images (unless proper data
augmentation is done). In the future, we would like to address
this shortcoming by systematically augmenting the dataset
or collecting more dataset from the practitioners to ensure
broader conclusions for the subject scenario.
R EFERENCES
[1] A. H. Warriner and K. G. Saag, “Osteoporosis diagnosis and medical
treatment,” Orthopedic Clinics, vol. 44, no. 2, pp. 125–135, 2013.
578
[2] J. Hays, J. K. Ockene, R. L. Brunner, J. M. Kotchen, J. E. Manson,
R. E. Patterson, A. K. Aragaki, S. A. Shumaker, R. G. Brzyski, A. Z.
LaCroix et al., “Effects of estrogen plus progestin on health-related
quality of life,” New England Journal of Medicine, vol. 348, no. 19,
pp. 1839–1854, 2003.
[3] L. Dalle Carbonare and S. Giannini, “Bone microarchitecture as an
important determinant of bone strength,” Journal of endocrinological
investigation, vol. 27, no. 1, pp. 99–105, 2004.
[4] G. M. Blake and I. Fogelman, “An update on dual-energy x-ray
absorptiometry,” in Seminars in nuclear medicine, vol. 40, no. 1.
Elsevier, 2010, pp. 62–73.
[5] S. R. Cummings and D. Black, “Bone mass measurements and risk
of fracturein caucasian women: A review of findings from prospective
studies,” The American journal of medicine, vol. 98, no. 2, pp. 24S–
28S, 1995.
[6] S. Schuit, M. Van der Klift, A. Weel, C. De Laet, H. Burger,
E. Seeman, A. Hofman, A. Uitterlinden, J. Van Leeuwen, and H. Pols,
“Fracture incidence and association with bone mineral density in
elderly men and women: the rotterdam study,” Bone, vol. 34, no. 1,
pp. 195–202, 2004.
[7] E. Lespessailles, C. Gadois, I. Kousignian, J. Neveu, P. Fardellone,
S. Kolta, C. Roux, J. Do-Huu, and C. Benhamou, “Clinical interest of
bone texture analysis in osteoporosis: a case control multicenter study,”
Osteoporosis international, vol. 19, no. 7, pp. 1019–1028, 2008.
[8] L. Pothuaud, E. Lespessailles, R. Harba, R. Jennane, V. Royant,
E. Eynard, and C.-L. Benhamou, “Fractal analysis of trabecular
bone texture on radiographs: discriminant value in postmenopausal
osteoporosis,” Osteoporosis International, vol. 8, no. 6, pp. 618–626,
1998.
[9] A. Taleb-Ahmed, P. Dubois, and E. Duquenoy, “Analysis methods
of ct-scan images for the characterization of the bone texture: First
results,” Pattern recognition letters, vol. 24, no. 12, pp. 1971–1982,
2003.
[10] T.-H. Hou and M.-D. Pern, “A computer vision-based shape-
classification system usingimage projection and a neural network,” The
International Journal of Advanced Manufacturing Technology, vol. 15,
no. 11, pp. 843–850, 1999.
[11] F. Mallard, B. Bouvard, P. Mercier, P. Bizot, P. Cronier, and D. Chap-
pard, “Trabecular microarchitecture in established osteoporosis: re-
lationship between vertebrae, distal radius and calcaneus by x-ray
imaging texture analysis,” Orthopaedics & Traumatology: Surgery &
Research, vol. 99, no. 1, pp. 52–59, 2013.
[12] R. Jennane, R. Harba, G. Lemineur, S. Bretteil, A. Estrade, and
C. L. Benhamou, “Estimation of the 3d self-similarity parameter of
trabecular bone from its 2d projection,” Medical Image Analysis,
vol. 11, no. 1, pp. 91–98, 2007.
[13] T. Lundahl, W. J. Ohley, S. M. Kay, and R. Siffert, “Fractional
brownian motion: A maximum likelihood estimator and its application
to image texture,” IEEE Transactions on medical imaging, vol. 5, no. 3,
pp. 152–161, 1986.
[14] A. Zaia, R. Eleonori, P. Maponi, R. Rossi, and R. Murri, “Medical
imaging and osteoporosis: Fractal’s lacunarity analysis of trabecular
bone in mr images,” in Computer-Based Medical Systems, 2005.
Proceedings. 18th IEEE Symposium on. IEEE, 2005, pp. 3–8.
[15] J. Pramudito, S. Soegijoko, T. Mengko, F. Muchtadi, and R. Wachjudi,
“Trabecular pattern analysis of proximal femur radiographs for osteo-
porosis detection,” Journal of Biomedical & Pharmaceutical Engineer-
ing, vol. 1, no. 1, pp. 45–51, 2007.
[16] S. Sevestre-Ghalila, A. Benazza-Benyahia, A. Ricordeau, N. Mellouli,
C. Chappard, and C. L. Benhamou, “Texture image analysis for
osteoporosis detection with morphological tools,” in International
Conference on Medical Image Computing and Computer-Assisted
Intervention. Springer, 2004, pp. 87–94.
[17] L. Houam, A. Hafiane, A. Boukrouche, E. Lespessailles, and R. Jen-
nane, “One dimensional local binary pattern for bone texture char-
acterization,” Pattern Analysis and Applications, vol. 17, no. 1, pp.
179–193, 2014.
[18] H. Oulhaj, M. Rziza, A. Amine, H. Toumi, E. Lespessailles, R. Jen-
nane, and M. El Hassouni, “Trabecular bone characterization using
circular parametric models,” Biomedical Signal Processing and Con-
trol, vol. 33, pp. 411–421, 2017.
[19] K. Zheng and S. Makrogiannis, “Bone texture characterization for
osteoporosis diagnosis using digital radiography,” in Engineering in
Medicine and Biology Society (EMBC), 2016 IEEE 38th Annual
International Conference of the. IEEE, 2016, pp. 1034–1037.
[20] L. Houam, A. Hafiane, R. Jennane, A. Boukrouche, and E. Lespes-
sailles, “Trabecular bone texture classification using 1d lbp and
wavelet coefficients in high-pass bands,” in International Conference
on Signal, Image, Vision and their Applications SIVA, vol. 11, 2011,
pp. 21–24.
[21] D. A. Pollen and S. F. Ronner, “Visual cortical neurons as localized
spatial frequency filters,” IEEE Transactions on Systems, Man, and
Cybernetics, no. 5, pp. 907–916, 1983.
[22] T. Ojala, M. Pietikainen, and T. Maenpaa, “Multiresolution gray-scale
and rotation invariant texture classification with local binary patterns,”
IEEE Transactions on pattern analysis and machine intelligence,
vol. 24, no. 7, pp. 971–987, 2002.
[23] Y. Kaya, M. Uyar, R. Tekin, and S. Yıldırım, “1d-local binary pattern
based feature extraction for classification of epileptic eeg signals,”
Applied Mathematics and Computation, vol. 243, pp. 209–219, 2014.
[24] K. Fukunaga, Introduction to statistical pattern recognition. Elsevier,
2013.
[25] V. Vapnik, The nature of statistical learning theory. Springer science
& business media, 2013.
[26] B. S. Manjunath, J.-R. Ohm, V. V. Vasudevan, and A. Yamada, “Color
and texture descriptors,” IEEE Transactions on circuits and systems
for video technology, vol. 11, no. 6, pp. 703–715, 2001.

579